Asynchronous Compute

Yeah still waiting on their solution for that, SeanP seems pretty confident about it though so there's definitely something in the works.

 

Who's SeanP? :nerd:

 

Seconded! Do tell! Is he an insider?

I hope it works out, because I'm tired of the ADFers rubbing it in our faces..

 

OK I'm sure Sean P. is Sean Pelletier, a senior product manager for NVidia..

Spets is probably referring to the tweet made by Sean P. when he said that asynchronous compute wasn't enabled in the drivers after Oxide released the beta 2 version of AotS:

Twitter feed.

 

Sorry but nvidia perf its not fine, only if you have a 980 ti maybe.

Look at the 970 in DX12

http://forums.guru3d.com/showpost.php?p=5243962&postcount=980

 

Yup, Sean Pelletier.
He's been pretty confident in some of his posts and tweets over the matter.
All in all after seeing TR and GoW DX12 performance on NV cards, I'm pretty sure it'll keep on the positive side of things. I mean every DX12 demo displayed by Microsoft has been on Nvidia GPU's.
Looking forward to Quantum Break performance

 

I get a nice fps now in AOTS, i think it is not Asyc but just a game update, Crazy avg 10 fps more now, 42 vs 56. 14 fps more after a game update?

 

AotS is rather badly optimized for NV h/w so it's no wonder that you may get +25% of performance with an update in it.

 

We will see

 

Which just goes to show there's a long way to go on the programming side. It's still early days people. I'm sure both camps will be happy once everything is ironed-out.

 

If the Maxwell NVIDIA graphics cards can do asynchronous compute then why don't the drivers support it? It's not like DX12 is new and I'm sure the driver team must have had ample opportunity to play around with DX12 code and games. I recall both Ashes of the Singularity and (the poorly optimized) ARK: Survival Evolved both being delayed because of problems with the NVIDIA drivers, prompting claims that "we are working with NVIDIA blah blah blah!".

Seems to me that NVIDIA do not seem very interested in showcasing their cards ability to handle DX12 but given the disappointing benchmark results so far I not hard to see why. In the last week, we've had three DX12 releases and two of those (Gears of War Ultimate Edition and Hitman) have issues. Rise of the Tomb Raider is the only one that runs without problems but the performance improvement does not seem that great at least from the benchmark.

I fear that Quantum Break will be another DX12 disappointment (some would say it already is by virtue of being released via the Windows Store only)... :3eyes:

 

I guess it's somewhere close to Nvidia not getting much of an improvement with dx12 over dx11 (unlike AMD), so not much use now. My personal guess is that they are a little worried how they will try to stand up against AMD under dx12 even though they have architectural differences that were great for dx11, but prove to be lacking a bit for dx12. At least for now I don't see Nvidia catching up to AMD just via drivers with Maxwell / older architectures.

 

Maxwell was not meant for dx12. Nvidia has their own strategy, sell now, care less in the future. They will market pascal as a dx12 ready selling point, and will sell tons, just like in every gpu launch. Thats their strategy and works, nvidia has a lot of market share.

AMD made this whole dx12 different, and their older cards benefit from it. That doesn't mean that r9 200/300 series will have much of a life, when polaris launches its all gpu's, from high-end to low, all of them will be powerful enough to eradicate r9 200/300 series, because they will become low end cards.

 

Which is why "future-proof" is meaningless in the PC world. Buy to enjoy today.

 

Asynchronous compute is a feature of DX12 API, every DX12 driver support it, even Fermi drivers will when they'll be released.

What you're saying is that NV's drivers do not support the specific way of running asynchronous compute on NV's h/w known as the way GCN h/w do this. This questions makes no sense however because there are no reasons why every h/w on the planet must run asynchronous compute specifically as GCN do this. Even more, there are no h/w on the planet presently which runs asynchronous compute in the same way as GCN h/w at all.

I really wish people would stop spreading AMD's FUD about async compute. ALL DX12 h/w support async compute because otherwise it wouldn't be DX12 compatible. There are different ways of running it on the h/w though and the way AMD does it isn't "a proper way" anywhere but on AMD's GCN h/w only.

You remember incorrectly. AotS was never delayed b/c of NV drivers and Ark was delayed b/c of the general state of DX12 support in UE4.

DX12 drivers are kinda bad for both NV and AMD at the moment. Which shouldn't surprise anyone really as this is just the first year of the new API being operational.

Two of those have issues on AMD's DX12 drivers.

Depends on what's a "DX12 disappointment". According to Remedy QB should be on the same level as in DX11 when running in DX12 on Maxwell and GCN.

 

It's more like DX12 was made for GCN, not Maxwell wasn't meant for DX12. You have to remember that DX12 is the API on Xbox One which features a GCN based GPU. Maxwell does support Async Compute, just not the definition that DX12/AMD specifies. Maxwell can do concurrent graphics/compute via CUDA/HyperQ, but it's not compatible with DX12's API being used in games. Whether or not they can engineer around that via drivers is up in the air. I personally don't think they can.

Regardless I'm not even sure how much DX12's Async Compute would help Nvidia's architecture. People seem to think it would just mirror AMD's, but that might not be the case. Async Compute is essentially acting like SMT in processors. If Nvidia is capable of keeping its pipeline filled more often than AMD is, which looks like the case in DX11, then Async Compute is going to do far less for Nvidia than it has for AMD.

Anyway, the bottom line is that AMD's effort getting into consoles and with Mantle was smart. Their cards were over-engineered compared to Nvidia's and that's starting to pay off for them. They played the long game and it's working now. The question is whether most users will remember this or even care for the next generation of cards. For me, personally, I upgrade almost every year, sometimes every two years. My 980 is now being outclassed by cards that are much cheaper, but I had always planned on dropping this thing anyway. So yeah I lose out on a few games that came out recently, but since 2014 I've enjoyed better performance than the AMD counterparts and now I'm dropping it. So the idea that AMD's cards are suddenly better doesn't really mean anything to me.

That being said, I'm thinking of picking up a Polaris 10, if it has 8GB of HBM.

 

Async compute is not the only thing for dx 12 and Async is no dx12 feature as well, its excist from 2004 and it is a Nvidia cuda thing. Go source voor Async compute you will see thats a 12 years Nvidia cuda thing.

Kernel calls are asynchronous from the point of view of the CPU so if you call 2 kernels in succession the second one will be called without waiting for the first one to finish. It only means that the control returns to the CPU immediately.

On the GPU side, if you haven't specified different streams to execute the kernel they will be executed by the order they were called (if you don't specify a stream they both go to the default stream and are executed serially). Only after the first kernel is finished the second one will execute.

This behavior is valid for devices with compute capability 2.x which support concurrent kernel execution. On the other devices even though kernel calls are still asynchronous the kernel execution is always sequential.

Check the CUDA C programming guide on section 3.2.5 which every CUDA programmer should read.

 

Multi-engine (yeah, the correct name for async-compute and async-copy) is a key-fature and a key-design of the entire Direct3D 12 API.

 

Stop ruining the circlejerk, NVIDIA's implementation is fine and they have had all this for more than a decade and there is just the mystery of why they don't implement it now in a way that actually increases performance. Their hardware is ready and capable.

Spoiler

 

I'm assuming that haven't as there is little to no performance to be gained from using it.
Hitman isn't exactly selling it as a huge selling point and that is from a AMD Gaming Evolved title with an engine known to be CPU heavy.